Six AIs Put Low Odds on a Quantum Break of RSA or ECC by 2030, and Split on Everything Else

In October the Financial Times reported that many tech giants now bet quantum computers will beat classical ones by 2030, with possible effects on cryptocurrencies, financial services and pharma research (Michael Peel, 11 October 2026, read through the French translation published by Courrier International). We use that framing as a question and nothing else from the piece.

The question is vague. "Beat" could mean many things, and "effects" could be anything. So we turned it into seven dated, checkable items, gave six AIs the same fact dossier, and asked each for a probability, a reason and two measurable indicators per item. We asked each model three times, so that run-to-run noise could be seen. We also asked what would change in the world by 2030 and 2035, and what an ordinary person, a company and a Bitcoin holder should do.

This article makes no prediction of its own and gives no financial advice. Its subject is where six AIs agree, where they split, and why.

The protocol

On 6 November 2026 we compiled a dossier from secondary sources (press, blogs, trackers). It holds hardware results and roadmaps, published resource estimates for breaking public-key cryptography, NIST standards and dates, and Bitcoin exposure figures. Each fact carries a label: secondary, unverified, contradictory, or primary source named but not opened by us. No primary paper or company page was opened. The dossier states that no 2025-2026 "quantum advantage" claim is established and that sector benefits are promoter claims, not measured results.

Six models answered three times each:

• Run 1 through the API.
• Runs 2 and 3 through each vendor's public chat app, in a fresh conversation each time, pasting the identical prompt.

The chat apps may run a different model version, system prompt, temperature or tool configuration than the API, and we did not control them. Run 1 and runs 2-3 are therefore not strictly the same experiment. We report all three runs and use the median so that one odd run does not decide a figure.

Rules: use only the dossier for 2025-2026 facts; mark any other fact "from model memory, unverified"; separate promoter claims from measured results; derive each number item by item and avoid a common template. Only Perplexity can search the live web (through its API and its app).

The seven items:

• M1. By 31 Dec 2028, a quantum advantage on a practically useful problem is accepted by independent peers: replicated, or not credibly beaten classically within 12 months.
• M2. By 31 Dec 2030, a machine runs a meaningful error-corrected algorithm on at least 100 logical qubits.
• M3a. By 31 Dec 2030, a quantum computer publicly breaks a real ECC-256 key or RSA-2048 modulus.
• M3b. The same, by 31 Dec 2035.
• M4. By 31 Dec 2030, most of the 20 largest banks have migrated external-facing systems to NIST post-quantum standards.
• M5. By 31 Dec 2030, Bitcoin has activated a quantum-resistant output type and at least 25% of exposed coins have moved to it.
• M6. The "quantum beats classical by 2030" framing: how plausible, with an explicit definition of "beats".

What the six answered

Median of three runs, with the range of the three runs in brackets. Probability in %. The last two columns give the median of the six model medians and the range of those medians.

The medians and ranges are our arithmetic. They give a sense of the middle and hide part of the spread. Two things are visible. Between models, medians differ by up to ninefold (M5, 3% to 28%). Within a model, the three runs often span a range as wide as the gap between models.

How stable is each model?

The widest within-model ranges:

• ChatGPT: M2 from 22 to 62, M4 from 20 to 55.
• Mistral: M4 from 10 to 60, M1 from 20 to 35, M5 from 3 to 20.
• Perplexity: M5 from 4 to 36, M4 from 8 to 43.
• Gemini: M3b from 35 to 68, M1 from 32 to 45.

For ChatGPT, Mistral and Perplexity, the API run was the highest or equal-highest on M4 and M5 (and on M2 for ChatGPT and Perplexity). The chat apps may differ from the API in model, mode and tools, so we cannot say whether this is run noise or a systematic difference between the two interfaces. Claude and Grok were the most stable across runs, with ranges of at most 20 points on any item. Stability is not accuracy: a model can be consistently wrong.

First observation: the cryptography break is the one consensus

Every model median puts a real ECC-256 or RSA-2048 break by 2030 at 5% or below, a narrow band next to every other row. Across all 18 runs the highest figure is 12% (Gemini, third run); 16 of 18 runs are at 5% or below. The shared reason comes straight from the dossier: the dossier's lowest estimate (under 500,000 physical qubits for ECC-256, attributed to a Google Quantum AI whitepaper of 31 March 2026, with circuits withheld and idealized assumptions; we did not open it) implies a gap of roughly three orders of magnitude to the machine sizes the models cite (hundreds to about 1,500 physical qubits, themselves unverified). The "1,000x" is our arithmetic on secondary figures, not a measured fact.

On ordering, the numbers are mostly coherent: a break by 2035 is never lower than by 2030 in any model's medians. One exception, in M1 and M6: "beats classical by 2030" should be at least as likely as "useful advantage by 2028" under a common definition, yet Claude's medians are the other way round (M1 22%, M6 15%), and this held in two of its three runs. The definitions may differ (see the third observation), but it is a point to flag. The coherence elsewhere comes from the question structure and the shared input. It does not show that any of the six knows more.

By 2035 the picture opens up. Claude's median is 6% and Gemini's 38%. In the first (API) run Mistral gave 40%, reasoning that NIST's 2035 deprecation date is "aligned" with fault-tolerance targets. A NIST disallowance date is a policy schedule, not a hardware forecast, so we count that as weak support. Mistral's median over three runs is 15%. Claude points to the 1,000x scaling still undemonstrated and notes a break could be done in secret or only on reduced key sizes.

Second observation: one fact, many readings

The sharpest divergence is M2, 100 logical qubits by 2030. The dossier says Quantinuum's Helios shows 48 logical qubits on 98 physical (secondary) and that vendors target fault tolerance around 2029-2030. All six cite this. In the first single-run version of this table, the answers split into two camps, ChatGPT and Gemini at 62-65% against Claude, Perplexity and Grok at 22-28%. Three runs blur that picture:

• High medians: Gemini 55%, Mistral 50%.
• Middle: Grok 35%, Claude 30%.
• Low medians: ChatGPT 25%, Perplexity 18%. ChatGPT's two chat runs gave 22% and 25%, against 62% in the API run.

So the split is not a stable two-camp divide. Only Gemini stays high in every run (45-65%) and Perplexity low in every run (12-24%). The others moved.

From the first (API) run, whose reasoning we read in full:

• Gemini called 100 logical qubits "a plausible linear progression"; ChatGPT said reaching 100 is plausible and the harder part is a meaningful error-corrected algorithm.
• Claude showed its arithmetic: about 45% that the roadmap delivers, times about 50% that the algorithm counts as "meaningful". It also cited unvalidated logical error rates and roadmaps that often slip (marked as model memory, unverified).

The 48-logical-qubit figure comes from a single secondary source, and a related "10x error-correction record" from October 2026 is single-source and unverified. The models differ on how to extrapolate from a claim nobody has independently checked.

One inconsistency worth flagging from the first run: Grok wrote that late-2026 machines "remain at hundreds of physical qubits" in M2, and "roughly 1,000-1,500" in M3a. Gemini also used about 1,500. We did not verify either figure.

Banks and Bitcoin: where the dossier is silent

Two items rest on facts the dossier does not contain, and they are the ones where single runs were least reliable.

M4 (banks). The dossier has NIST's standards finalized in August 2024 and NIST IR 8547 planning to deprecate vulnerable algorithms after 2030 and disallow them after 2035. It has nothing on the banks. Model medians run from 8% (Perplexity) to 45% (Grok), and single runs from 7% to 60%. Claude's first-run reasoning was as much about resolution as about banks: "completed" for most of 20 global institutions is a high bar, and public proof of completion is unlikely. ChatGPT saw strong reasons to migrate and no bank-by-bank evidence. Same facts, different handling of the missing evidence.

M5 (Bitcoin). The dossier gives about 6.4 million BTC (32.2% of supply) "in outputs with exposed public keys" per the Quantum Alliance of nine firms (October 2026), about 6.9 million per European supervisory authorities, and "25-40%" per a tracker. These are secondary, differently defined and contradictory; the dossier does not say whether they count old P2PK outputs, reused addresses or keys revealed after spending, and we could not check. A proposal named BIP-360 is reported as merged into the BIPs repository on 11 February 2026; this comes from one source and we did not check the repository. Merged proposal text would not be activation. Model medians run from 3% (Claude) to 28% (Grok). Claude split M5 into two steps in its first run, about 12% for activation by 2030 and about 12% for 25% of exposed coins moving given activation. Perplexity moved from 36% in its API run to 6% and 4% in its chat runs. Both read the same community division on migration and on frozen or abandoned coins.

Third observation: M6 is not one question

M6 asks about the FT framing, and we asked each model to define "beats". In the first (API) run the definitions differed, so the six numbers were not answers to the same proposition. We later read the full chat-run answers for M6. The medians in the main table do mix several wordings per model, and some models gave two readings in one answer:

• Claude gave a strict reading and a weaker one (second run: 15% strict, about 45% for the weak reading); the table carries the figure it headlined.
• Mistral defined "beats" in two ways within its runs.
• Perplexity (third run) gave a two-sided definition: a broadly accepted, independently replicated advantage on a useful task is "weakly supported", a narrow simulation advantage is "higher but still not established". It put one number, 20%, on the stricter reading.
• ChatGPT defined useful and beats explicitly.
• Mistral's third run did not give "P = xx%" but wrote "~30% plausible", with the definition "independent replicated advantage on a useful problem"; we recorded 30.

The first-run definitions fall into two readings:

• Reading A: six related but non-identical definitions of a single useful advantage (first-run figures ChatGPT 57%, Gemini 55%, Mistral 45%, Perplexity 39%, Grok 31%, Claude 28%).
• Reading B, broad, sustained superiority: only Claude gave a figure, under 2%. The other five did not state one.

The definitions add different criteria (independent acceptance, commercial relevance, sustained advantage, replication, economic value), so the 32.5% median in the main table is a median of related but different propositions, not of one. Treat it as indicative only.

So the headline number answers a narrow question: one useful task, once. No model gave a high probability to the broad reading, though only Claude put a number on it. A headline that says "quantum beats classical by 2030" can be roughly true or clearly false depending on whether "beats" means one niche result or general superiority. The dossier adds that 2025-2026 advantage claims remain contested or unsettled, and the IBM claim of July 2026 rests on preprints and is called a "candidate" by trackers.

What actually changes in the world

We asked for a view by sector, for 2030 and 2035, with confidence labels, separating promoter claims from measured results. The table below is built from the first (API) run. We later read the sector answers of runs 2 and 3 on screen. They did not change the picture, but two formatting facts matter: Gemini's and Perplexity's Q7 were templated, one "2030: ... 2035: ..." line per sector, which makes the sector lines look more parallel than the evidence is; and the models differ on whether they name the harvest-now-decrypt-later risk (see "What these numbers do not show"). The consensus is thin and mostly negative: the models can name few measured changes.

See the protocol-violations section below for out-of-dossier claims; we do not count them as findings. In the first run Gemini said export controls on cryogenics will follow a "quantum race", and that data-centre energy for some AI workloads will fall measurably by 2035. Neither has support in the dossier, and the second appears with no source. Gemini also wrote "high confidence in quantum-assisted catalyst design" for 2035 and tagged it "from model memory, unverified", which contradicts itself: unverified memory does not justify high confidence.

The takeaway is simple. The six agree that cryptography is the one place where a decision has to be made now, because the NIST dates (2030 and 2035) and the shelf life of secrets matter more than the date of any machine. Everything else is promoter territory until independent results exist.

What the panel said to do

Practical advice, summarized from the six first-run answers, with no investment advice and no price targets.

• An ordinary person: keep devices, browsers and apps updated, since post-quantum protection arrives through updates. Do not buy "quantum-proof" products or schemes. Do not panic over claims that encryption is already broken. (ChatGPT, Claude, Perplexity, Gemini.)
• A company: inventory where RSA and elliptic-curve cryptography are used, prioritise data that must stay confidential for ten years or more, test NIST ML-KEM and ML-DSA libraries in staging, set an owner for migration, and ask vendors for timelines. Claude adds NSA's CNSA 2.0 milestone of January 2027, which comes from the dossier as a secondary source. Be sceptical of quantum "advantage" services until independently tested.
• A Bitcoin holder: do not act on AI advice. Moving coins carries its own risks: fees, address errors, lost seeds, tax exposure and wallet compromise. Have a qualified Bitcoin security professional assess your wallet-specific exposure. For the record, the panel disagreed: Claude, Grok and Gemini leaned toward moving coins from reused or old exposed outputs to fresh addresses once wallet support is verified; ChatGPT and Perplexity warned against rushing and against treating a merged proposal as protection; Mistral suggested "cold storage", which would not hide a public key already on the chain. Nothing here is financial advice.

Protocol violations and sensitivity

The protocol asked models to use only the dossier for 2025-2026 facts and to flag anything else as unverified model memory. In the first run, Gemini broke this in its sector-impact answer: it wrote that BIP-360 is "likely active" in 2030, predicted export controls on cryogenics and a measurable data-centre energy reduction by 2035, and gave "high confidence" in quantum-assisted catalyst design while tagging it "from model memory, unverified". We exclude those statements from the sector table and from our conclusions. We then read runs 2 and 3 for the same problems. Findings: Perplexity used live web search in its two quantum chat runs (11 to 12 web sources shown) and said in its Limits line that NIST IR 8547 details were checked against NIST's publication, a deliberate step outside the dossier that is not marked "model memory" in the body of the answer; its Q10 gave two confidence levels (Medium for the cryptographic timeline, Low for hardware scaling) where one was requested; Grok's quantum answers were very short, four items were identical across its runs, and it flagged no model-memory facts; Claude flagged its out-of-dossier points. Gemini's probabilities for M1-M6 are kept in the tables because they are stated from dossier facts, but they are the highest or among the highest on several items, so here is the median of the model medians without Gemini:

With three runs the dependence on Gemini is smaller than in the single-run table, where the M2 median fell by 11 points without Gemini. Now every median moves by 7 points or less (largest: M1, 28.5 to 22). Gemini still sits at the top of the distribution on M1, M2, M3a, M3b and M6, so the panel median is slightly higher with it. This is a sensitivity check, not a verdict on which version is right.

What each model said it was least sure of

From the first (API) run:

• ChatGPT: M3b, M4 and M5.
• Claude: M2 and M6 (definition-dependent) and M5 (a joint estimate).
• Gemini: M3b and M5, because they depend on human governance and long-term scaling.
• Grok: M1 and M3a, because they extrapolate across three orders of magnitude without intermediate data points.
• Mistral and Perplexity: M3a and M5 (plus M3b for Perplexity).

The runs partly confirm this. The items where models moved most were M4 and M5 (ChatGPT, Mistral, Perplexity), M2 (ChatGPT) and M3b (Gemini). M3a, flagged as shaky by several models, was in fact the most stable item across runs.

The missing facts they asked for converge: independently validated logical-qubit counts and error rates (several models); peer review and classical-rebuttal status of the IBM and Google advantage claims (several models); reproducible resource estimates for breaking ECC-256 and RSA-2048 (several models); the governance and activation status of BIP-360 (several models); bank migration inventories (several models).

What these numbers do not show

Agreement is not validation. The six models read one dossier. Their low odds on a 2030 key break come from one resource estimate repeated six times, so the convergence measures how strongly the dossier frames the question, not whether the estimate is right. "All six agree" is a statement about the panel, never evidence about the world.

The gap can close without new hardware. The "three orders of magnitude" is a hardware comparison. The dossier itself shows resource estimates for RSA-2048 falling from about 20 million physical qubits (2019) to under 1 million (2025) and, in one simulation-only claim, under 100,000 (2026), all on secondary pages we did not verify. Algorithmic progress can shrink the requirement as well as chip progress can grow the machine. The panel's numbers mostly treat the gap as a hardware question, and we did not ask the models to separate the two.

Harvest-now-decrypt-later is the actionable risk. Encrypted data copied today can be read later if a break arrives, so the risk to long-lived secrets does not wait for 2030. The concept is not in the dossier. After reading all runs: Claude, Mistral and Perplexity named it (Perplexity under cybersecurity 2030 in both chat runs), Gemini only in its first run, in the crypto line and the "store now, decrypt later" advice, and Grok did not. It underlies the "ten years or more" advice. We label it model memory, unverified. It is the one item in this article that calls for action whatever the date of a break.

Some items are poorly calibrated questions. M4 ("most of the 20 largest banks have migrated") does not say whether a pilot, a hybrid deployment, full production or a certification counts. The 8% to 45% spread may reflect these readings as much as a real disagreement about banks, and we cannot separate the two. M5 is a double condition, activation and 25% of exposed coins moved, which pushes the figure down mechanically; Claude's two-step split in its first run (about 12% times 12%) shows how. A cleaner design would ask the two parts separately.

Why a model moves is an open question. Claude and Grok were the steadiest across runs; ChatGPT, Mistral and Perplexity moved most between the API run and the chat runs. A model's stability could reflect the quality of its reasoning, or only its sampling settings and interface mode. With three runs and uncontrolled chat settings we cannot tell which, and we do not claim to. Reading the reasoning adds one clue, not a proof: Gemini's large same-model gaps (M3b 38 against 68 across runs) came with a different Q10 confidence label from run to run, while Grok's stability partly reflects answers that were short and near-identical across runs, which is repetition more than reasoning.

Predictive value is untested. We show that noise exists. We do not show that the signal does. Settling it would need many more runs, a comparison with human expert forecasts, or a backtest on questions that have already resolved. We have done none of these.

Weighting. Models are not weighted by protocol compliance or stability. Gemini, which left the dossier in its first-run sector answers, counts in the medians like the others; the table above is the only treatment of that.

The practical rule

Use an AI to find out which facts carry which answer, not to get odds.

1. Look for the shared input. The 5% ceiling on a 2030 key break comes from one dossier fact repeated six times. It is not six independent confirmations.
2. Never trust one run. One AI answer is one draw from a distribution. Here a model's range across its own three runs was often as wide as the gap between models (ChatGPT on M2, Perplexity on M5).
3. Read the spread. When two models differ widely on the same fact, ask which assumption differs: the extrapolation (M2), the missing evidence (M4, M5) or the definition (M6).
4. Do the cryptography now. The decision that does not depend on any forecast is the inventory of public-key use. It costs the same whether the first useful machine arrives in 2029 or 2035.

Indicators to watch

Named by the models and kept for the update log:

• Logical qubits (M2): a peer-reviewed run of 100 or more logical qubits with stated logical error rates; independent verification of Quantinuum's reported 48 logical qubits and 10x error-correction claim.
• Advantage (M1, M6): replication of a useful result on independent hardware; an IBM advantage claim still unrefuted 12 months after posting; an audited end-to-end cost comparison against tuned classical baselines.
• Cryptanalysis (M3a, M3b): independent validation of the sub-500,000-qubit ECC-256 estimate with circuits released; a machine above 100,000 high-fidelity physical qubits.
• Banks (M4): at least 10 of the top 20 publishing dated completion for internet-facing endpoints; a binding regulator deadline before 2030.
• Bitcoin (M5): activation code released and reaching its signalling threshold; on-chain migration above 5% of exposed supply in the first year.

Method and limits

• Six models answered the same dossier and the same questions on 6 November 2026, three times each, with identical rules: run 1 through the API, runs 2 and 3 through the vendors' chat apps in a fresh conversation. For run 1, the Gemini answer was first returned cut off in the middle of the third item; we re-ran that one model once and use the complete answer. The first, truncated output is archived.
• The chat apps may use a different model version, mode, system prompt or tool configuration than the API. We recorded the mode each app displayed (table above) and nothing more. Perplexity's quantum chat runs used live web search (11 to 12 sources shown) and cite NIST material outside the dossier; its Mideast chat runs showed no sources.
• The three runs are not independent draws from one fixed setup. We use the median and the range, and we do not compute confidence intervals on three points.
• Chat apps have conversation memory on some accounts. We used a fresh conversation each time but cannot rule out account-level personalization.
• The probabilities for runs 2 and 3 were first extracted from the on-screen answers by pattern. Afterwards we re-opened all 12 quantum chats (6 models, 2 runs) and read each answer on screen, scrolling from the top, in screenshots rather than copied text, and compared the probabilities with the stored values: they matched in every chat read. Reading was by screenshot, so it is a human-style read, not a text diff. One run (Mistral, third) did not use the requested "P = xx%" format.
• During the chat runs the tester's shared clipboard was overwritten once, and one unrelated message was sent by mistake into a Claude conversation. That conversation is not part of the data; each retained run was checked to hold the correct prompt before sending, and its answer to contain the seven expected items.
• The dossier is secondary-source only. We read the opening of the Courrier International translation of the FT piece (paywalled), and nothing more of it is used or attributed to it. No peer-reviewed validation was checked.
• Probabilities are model opinions on one date. They are not forecasts, not statistics and not calibrated against outcomes.
• Medians, ranges and the ordering checks are our arithmetic on the stated figures.
• A model's stated training cutoff limits what it can add beyond the dossier: Mistral says its cutoff is 2023. Facts from outside the dossier are used only when the model flagged them.
• No investment advice, no price predictions. English draft only; not published, not submitted to search engines.

Update log

This article is built to be updated. When new verified facts arrive, we will rerun the same seven items on an updated dossier and append the figures, keeping every earlier row.